Ultralytics YOLO26: retail shelf detection tutorial
Learn to train YOLO26 on SKU-110K, inspect shelf detections, and export models for Intel, Nvidia or Apple hardware.
Player not loading? Watch on YouTube
This tutorial uses Ultralytics YOLO26 and the Ultralytics Platform to train a detector on SKU-110K, described by the presenter as roughly 12,000 retail shelf images with around 8,000 training images. The dataset has one object class. The detector locates items rather than identifying individual products; shelf gaps and restocking alerts require additional application logic.
The walkthrough selects a nano model and 50 training epochs. It demonstrates cloud training and shows where to obtain a command for local training on your own machine. The presenter then reviews precision, recall, mean average precision and loss curves before testing predictions with sample images. The reported example takes about 140 milliseconds for inference, with roughly 500 milliseconds of network latency. These are observations from the demonstration, not general performance guarantees.
For local AI deployment, the export section covers downloading the PyTorch model and converting it to OpenVINO for Intel hardware, TensorRT for Nvidia, or Core ML for Apple hardware. Export settings include image size, batch size and precision. The presenter claims FP16 can improve speed without accuracy loss and notes that INT8 may reduce accuracy. The tutorial also discusses edge inference for frequent images or video, while suggesting occasional cloud requests for less frequent shelf checks.